Resource Management Algorithms for Computing Hardware Design and Operations: From Circuits to Systems

Abstract

The complexity of computation hardware has increased at an unprecedented rate for the last few decades. On the computer chip level, we have entered the era of multi/many-core processors made of billions of transistors. With transistor budget of this scale, many functions are integrated into a single chip. As such, chips today consist of many heterogeneous cores with intensive interaction among these cores. On the circuit level, with the end of Dennard scaling, continuously shrinking process technology has imposed a grand challenge on power density. The variation of circuit further exacerbated the problem by consuming a substantial time margin. On the system level, the rise of Warehouse Scale Computers and Data Centers have put resource management into new perspective. The ability of dynamically provision computation resource in these gigantic systems is crucial to their performance. In this thesis, three different resource management algorithms are discussed. The first algorithm assigns adaptivity resource to circuit blocks with a constraint on the overhead. The adaptivity improves resilience of the circuit to variation in a cost-effective way. The second algorithm manages the link bandwidth resource in application specific Networks-on-Chip. Quality-of-Service is guaranteed for time-critical traffic in the algorithm with an emphasis on power. The third algorithm manages the computation resource of the data center with precaution on the ill states of the system. Q-learning is employed to meet the dynamic nature of the system and Linear Temporal Logic is leveraged as a tool to describe temporal constraints. All three algorithms are evaluated by various experiments. The experimental results are compared to several previous work and show the advantage of our methods.